Navigating the labyrinth: How generative models tackle complex data sampling